Artificial intelligence technology has reached a point where it is disrupting many fields, including photography.
AI-infused photography and photo-editing programs are wither now available on many mobile devices, or soon will be. Rather than focus on adding more megapixels, the most fascinating innovations in smartphone photography are because of machine learning algorithms, which are making our photos and selfies better than ever. In a recent blog post, Google showed how its AI-powered photo editing technology was able to edit photos to the point that professional photographers rated them as professional or semi-pro quality.
One of the most striking examples of AI in photography right now can be seen in 'portrait mode'. For the last few years, some phones use dual rear cameras to detect depth. Then, state-of-the-art machine learning is used to automatically blur the background, emphasizing the subject.
The result is AI technology producing some truly spectacular images, even for bad photographers. Experts have said the evolution of this technology could allow for a portrait mode function that uses only one camera, as machine learning algorithms will soon be able to pick the subject out of an image without the use of depth information from a second camera.
“Zoom and enhance”
If you’ve ever watched scripted police shows like CSI, you’ve probably see characters try to identify a suspect on a grainy video through a function called ‘zoom and enhance’ – where a program zooms in on a person in the video and magically makes their face visible.
Until recently, this function was impossible because a computer cannot get more detail than what was in the original source material. But now, Google has developed an AI system that can recreate facial images from highly-pixelated source material. The system uses one neural network to map the source image to low-quality versions of other, similar high-resolution imagery and another neural network adds details to the source image by analysing a large volume of other similar images.
While the results of Google’s system aren’t perfect, its enhanced images of celebrity faces have compared favourably to regular non-upscaled, hi-res images of those same faces.
Anyone who has tried to erase a tattoo or move a person across an image using Photoshop knows it can be a painstaking process. According to Adobe, it is currently working on an AI-powered version of Photoshop that can pick out objects in an image and allow a user to edit them with basic clicks and drags.
The company is also working on using AI to allow users to improve the quality of pictures taken with their smartphone camera. Smartphone cameras use a wide-angle lens that is great for landscape shots but can cause unflattering distortions in portraits. According to Adobe, an AI system currently in development will allow users to modify their images so that they appear to have been taken with a telephoto lens.
Fake photos have been around as long as photography and AI is going to make fakes even more convincing.
Generative adversarial networks, or GANs, are a new type of neural network that can quickly produce fake images based on machine learning and a massive database of images. A GAN is comprised of two networks, a “generator” and a “discriminator”.
The generator studies real images in order to learn how to make more convincing fakes. Fake images produced by the generator are then passed to the discriminator, which rates each fake based on how realistic it appears. As time passes, the generator “learns” how to make more and more convincing fakes, while the discriminator gets better at sniffing out fake images.
While the prospect of more-convincing fake imagery seems like an ominous development for the rise of ‘fake news’, it would mean a more realistic experience for video game players and movie goers.